Performance Summary#

This document provides performance benchmarks for various large language models using NeMo Automodel with the PyTorch backend.

Pre-Training Performance#

The table below shows training performance for full sequences with no padding across different model architectures and scales.

Model	#GPUs	GBS	MBS	LBS	GA	Seq Length	TP	PP	CP	EP	VP	FSDP	Kernel Optimizations	Time per Global Step (s)	Model TFLOPs/sec/GPU	Tokens/sec/GPU
Nemotron V3 Nano 30B (26.02)	8	512	4	4	16	4096	1	1	1	8	-	8	TE + DeepEP + TorchSDPA	15.614	328	16,789
DeepSeek V3 671B	1024	8192	1	8	4	4096	1	4	1	64	8	256	TE + DeepEP	37.87	216	865
DeepSeek V3 671B	256	512	1	8	1	4096	1	4	1	64	8	64	TE + DeepEP	8.18	250	1,002
Kimi K2	256	512	1	8	2	4096	1	8	1	32	4	32	TE + DeepEP	8.86	189	924
Qwen3 MoE 30B	8	512	4	4	16	4096	1	1	1	8	-	8	TE + DeepEP	21.773	277	12,040
GPT-OSS 20B	8	256	2	2	16	4096	1	1	1	-	-	8	TE + DeepEP + FlexAttn	10.04	279	13,058
GPT-OSS 120B	64	512	2	2	4	4096	1	1	1	-	-	64	TE + DeepEP + FlexAttn	4.30	231	7,626

The table below shows finetuning (LoRA) performance for full sequences with no padding across different model architectures and scales.

Model	#GPUs	GBS	MBS	LBS	GA	Seq Length	TP	PP	CP	EP	VP	FSDP	Kernel Optimizations	Time per Global Step (s)	Model TFLOPs/sec/GPU	Tokens/sec/GPU
Llama3 8B	1	32	2	2	16	4096	1	1	1	-	1	1	-	10.51	402	12472.87
Qwen2.5 7B	1	32	2	2	16	4096	1	1	1	-	1	1	-	9.29	423	14110.05
Llama3 70B	8	32	1	4	4	4096	2	4	1	-	10	1	-	24.87	190	658.62
Qwen2.5 32B	8	32	1	8	2	4096	1	4	1	-	8	1	2	8.40	261	1950.93
Llama3 70B 2-node	16	32	1	4	2	4096	2	4	1	-	10	1	2	12.03	197	680.74
Qwen2.5 32B 2-node	16	32	1	8	1	4096	1	4	1	-	8	1	4	4.48	244	1826.49

MFU: Model FLOPs Utilization - ratio of achieved compute to peak hardware capability
TP: Tensor Parallelism - splits individual layers across GPUs
PP: Pipeline Parallelism - splits model layers into stages
EP: Expert Parallelism - distributes MoE experts across GPUs
DP: Data Parallelism - replicates model and splits data
VP: Virtual Pipeline - number of pipeline stages per GPU for interleaving
MBS: Micro-Batch Size - size of one forward pass in pipeline
LBS: Local Batch Size - size of one step per GPU
GBS: Global Batch Size - total batch size across all GPUs
GA: Gradient Accumulation - number of local-batches before optimizer step
TE: Transformer Engine kernel optimizations - RMSNorm, Linear and DotProductAttention
DeepEP: Deep Expert Parallelism - advanced EP routing for MoE models
FlexAttn: PyTorch’s Flex Attention

Pre-training benchmark configurations are available in examples/benchmark/configs/ and fine-tuning (LoRA) configurations are in examples/llm_finetune/:

deepseek_v3_te_deepep.yaml - DeepSeek V3 with TE + DeepEP
kimi_k2_te_deepep.yaml - Kimi K2 optimized configuration
qwen3_moe_30b_te_deepep.yaml - Qwen3 MoE with TE + DeepEP
gptoss_20b_te_deepep.yaml - GPT-OSS 20B with optimizations
gptoss_120b_te_deepep.yaml - GPT-OSS 120B optimized
llama3_1_8b_peft_benchmark.yaml - Llama-8B Finetuning (LoRA) optimized
qwen2_5_7b_peft_benchmark.yaml - Qwen2.5-7B Finetuning (LoRA) optimized
custom_llama3_3_70b_instruct_peft_benchmark.yaml - Llama-70B Finetuning (LoRA) optimized
qwen2_5_32b_peft_benchmark.yaml - Qwen2.5-32B Finetuning (LoRA) optimized
custom_llama3_3_70b_instruct_peft_benchmark_2nodes.yaml - Llama-70B Finetuning (LoRA) optimized on 2 nodes
qwen2_5_32b_peft_benchmark_2nodes.yaml - Qwen2.5-32B Finetuning (LoRA) optimized on 2 nodes

Note

All benchmarks use mock data for consistent performance measurement.
Fake balanced gate is enabled to simulate ideal expert routing.
No gradient clipping applied for pure performance measurement.
MFU calculated using peak TFLOPs for the system (989 for BF16 H100).
Step times include forward and backward passes + optimizer step for the global batch.